Nacionalni portal odprte znanosti

In EM and related algorithms, E-step computations distribute easily, because data items are independent given parameters. For very large data sets, however, even storing all of the parameters in a single node for the M-step can be impractical. We present a framework which fully distributes the entir ...